The Rare Clean Stretch
Start with the thing everyone knows, because it is almost always true. Shine light into a material that absorbs, and the light dies exponentially. Every millimetre takes the same fixed fraction; the survivors halve, and halve, and halve. Beer and Lambert wrote it down two centuries ago and it has been quietly correct in ten thousand laboratories since: the amount of light left after depth L is e^(−αL), and α — how fast the dying goes — is set by how much absorbing stuff is in the way. More absorber, steeper death. It is the law behind sunglasses and stained glass and the darkness at the bottom of the sea.
So here is a fair question with an obvious answer. Take a clean waveguide — a channel that carries light with almost no loss — and start poisoning it. Drop in absorbing sites here and there, at random, sparsely: most of the channel stays perfect, but every so often there’s a site that eats light. As you add more poison, or make each poisoned site hungrier, the light should die faster. That’s Beer–Lambert. More loss, more absorbed, quicker dark.
Both halves of that obvious answer are wrong. Not wrong at the edges — wrong at the center. And the two ways they’re wrong turn out to be the same physics wearing two faces.
The first wrongness: the light won’t die on schedule
Watch a pulse spread through the poisoned channel and measure how much survives as time goes on. For a little while it behaves: the survival curve drops the way an exponential drops. But then it peels off. It refuses to keep halving. On a plot of survival against time it bends the wrong way — concave, hanging on — and at long times there is vastly more light left than any exponential would ever allow. Spread the same total amount of poison uniformly through the channel instead of clumping it into rare strong sites, and the light is gone almost instantly. Clumped and sparse, it lingers to a hundred times longer, a thousand.
The decay law isn’t exponential at all. It’s stretched: the logarithm of the survival grows not like t but like a fractional power of t, −ln P ~ t^β with β less than one. A stretched exponential is the signature of a system that has hiding places.
Where are they hiding? This is the whole answer, and it is worth saying slowly. The light that survives is the light that fell, by luck, into a rare long stretch of channel with no poison in it. Randomness sprinkled sparsely leaves gaps. Most gaps are short. But in a long enough channel there will be, somewhere, an unusually long clean run — twenty sites, forty, with no absorber at all — and a run like that is a little cavity, a room with soft walls, where light can rattle back and forth for a very long time before it leaks out the ends and finds an absorber. The typical light dies on Beer–Lambert’s schedule. The surviving light is not typical. It is the light that found the rare clean room, and it reports back a completely different law.
You can prove this isn’t a story. Take the actual random channel, decompose it into its clean runs, and rebuild the survival curve from two ingredients only: how often a clean run of each length occurs (pure combinatorics — long runs are exponentially rare), and how long a lone clean run of that length keeps light alive (one small simulation per length). Add them up, weighted by rarity. The rebuilt curve lands on the true one to within a few percent across the whole decay. Nothing but rare clean runs and their lifetimes. The mechanism is the answer.
The lineage: this is an old ghost in a new house
Physicists have met this ghost before, in a completely different material. In 1969 Robert Griffiths was studying a magnet with holes punched in it at random — a dilute ferromagnet — and proved something that unsettled people. Even in the temperature range where the diluted magnet as a whole refuses to order, its free energy is not a smooth, analytic function. There is a hidden non-smoothness, a singularity, spread across a whole band of conditions. The reason: rare regions. By chance, somewhere in the disorder, there is a patch with no holes — a chunk of pure magnet — and a large enough pure patch behaves locally like it has ordered, even when the bulk hasn’t. These rare ordered islands, exponentially unlikely but not impossible, leave their fingerprint on everything. The band where they rule is now called the Griffiths phase, and the fingerprint a Griffiths singularity.
The same skeleton holds up several other famous results. Ilya Lifshitz, in the early 1960s, showed that the density of quantum states right at a band edge of a disordered solid tails off not sharply but as a soft exponential of an inverse power of the energy — because the lowest, longest-lived states live in rare impurity-free regions. Donsker and Varadhan, and independently Balagurov and Vaks, proved the time version: a particle wandering among randomly placed traps survives, at long times, as a stretched exponential exp(−c t^(d/(d+2))) — in one dimension, t^(1/3) — and the survivors are precisely the particles lucky enough to be born in a large trap-free void, diffusing quietly while everyone else is caught. Rare regions; exponential rarity; anomalously long life; a stretched-exponential tail. It is one idea, and Longhi’s contribution is to find it living in light — in a photonic lattice you could actually build on a chip — and to show, carefully, that it is this and not some look-alike (not an exceptional point, not an interference fringe). The clean rooms are real, and the light finds them.
A twist the diffusive ghost didn’t have
Here the light does something the wandering particle didn’t, and it’s worth being honest that the exponent changes. For a diffusing particle the clean void’s survival is set by how long diffusion takes to reach the walls — the lowest note of a drum of length ℓ, a rate that falls off like 1/ℓ². Run the rarity-versus-longevity competition with that and you get Donsker and Varadhan’s one-third.
Light in a coherent lattice is not diffusing. Its long-lived clean-room mode leaks out through the tiny amplitude it has at the very edge of the room, right where the first absorber waits — a golden-rule leak — and that edge amplitude is smaller, so the rate falls off faster, like 1/ℓ³. The longer the clean run, the dramatically longer light lives in it: measured directly, across runs from sixteen sites to a hundred and twenty-eight, the lifetime stretches as the cube of the length, clean as a ruler. Put that longevity into the same competition — rare runs cost exp(−ρℓ) to find, and pay off with a lifetime that grows as ℓ³ — find the run length that best balances the two at each time, and the arithmetic hands you a different stretched exponent: β = 1/4. The coherent clean room is a better hiding place than the diffusive void, and the survival law knows it.
(An honest seam, because a proof-walk owes you its soft spots: one-quarter is the asymptotic answer, the law at infinite time. Over the times you can actually simulate, the measured exponent sits nearer a third and slides slowly downward toward a quarter as you watch longer — because the sparse absorbers aren’t perfect walls; at moderate strength a single lossy site passes half the light it meets, so neighboring clean rooms leak into each other and the effective rooms are larger than they look. The number you measure at finite time is a real crossover between the coherent quarter and something more diffusive, not a clean plateau. Reproducing the paper is watching that crossover, not declaring a tidy constant. The mechanism — rare clean rooms carry the tail — is exact; the last digit of the exponent is a limit you approach, and saying so is part of the work.)
The second wrongness: there is a best amount of loss
Now the other face, and it is the one that would surprise an engineer most. Forget survival over distance; ask simply how much of an incoming beam a single absorbing site swallows, as you turn its hunger up from nothing to enormous. Beer–Lambert says: monotone, more is more. The truth is a hill.
You can do this one exactly, in closed form, no simulation. Send a wave of the lattice at a single site whose loss strength is γ. The wave partly transmits, partly reflects, partly is eaten; conservation bookkeeping gives the eaten fraction as
A(γ) = 2vγ / (v + γ)²,
where v is the light’s speed through the clean lattice. That formula is a hill with a summit. At γ = 0 the site is transparent and eats nothing. As γ climbs, absorption climbs — for a while, as expected. It peaks exactly when the loss is matched to the light’s speed, γ★ = v, and at that summit it swallows precisely half. And then — this is the surprise — as you keep cranking γ up, absorption falls back toward zero. A very hungry site eats less.
Why? Because a strongly dissipative site stops being a mouth and becomes a wall. Turn the loss up high enough and the site is so lossy that light can no longer even enter it — it is reflected at the threshold, sent back or routed around, and light that never enters can never be absorbed. It is a cousin of the quantum Zeno effect, where watching a system hard enough freezes it: measure a site to death and you decouple it. The hungriest absorber is a mirror. So the way to absorb the most is not the most loss — it is just enough, matched, and no more. Less is more, with a formula under it.
And now the two faces meet, because they were always one. The sparse channel absorbs anomalously little at long times for the same reason a too-hungry site absorbs little: the light survives in the places the loss isn’t, and strong loss builds the very walls that keep light out of the loss. Rare clean rooms and reflecting mouths are the same sentence read forward and backward. To catch the light you must neither starve the absorbers nor gorge them, and you must not leave long clean rooms for the light to hide in. The material that absorbs best is not the one with the most darkness in it. It is the one with the darkness placed just so.
What is true, and what is only described
It would be easy to end on the pleasing inversion — more loss, less absorbed; add darkness and light survives — and stop. But the paper, and honesty, ask for a verdict, and the verdict is partial in the good way.
What is settled: the exponential intuition is simply wrong for sparse disorder, and it is wrong for a nameable reason that reproduces to a few percent from first principles — rare clean runs, exponentially unlikely and exponentially long-lived, carry a stretched-exponential tail, and the same rare-region logic that Griffiths found in a punctured magnet and Donsker and Varadhan in a field of traps runs the show here in light. And the single-site absorption hill, γ★ = v and a cap of one half, is exact algebra you can check in a line. These are not metaphors. They are the material’s actual behavior.
What is only described, not explained away: the precise stretched exponent is a limit approached, sensitive to how transparent the sparse absorbers are, and the whole picture is a disorder-averaged, long-time asymptotic statement — the kind of law that is exactly right about the ensemble and says nothing about which particular channel you hold in your hand. Longhi’s own care is the model here: he does not claim the light knows Griffiths’ theorem; he shows the fingerprint is present and rules out the imposters. The rare clean stretch is a description of where the surviving light goes. That it survives at all, longer than any honest exponential would permit, is the fact. Why the world so often hides its longest-lived things in its rarest empty places — that is older than any of these papers, and none of them close it.
Companion to the verified reconstruction in research/the-rare-clean-stretch/ (SPINE_MD5 394e5731…): uniform loss is exactly exponential; a single site has an exact absorption optimum at γ★=v with cap ½; sparse loss decays sub-exponentially, and the whole decay curve is rebuilt, to within a few percent, from run-length statistics times single-run longevity. Sources: Longhi, arXiv:2607.03205 (2026); Griffiths, Phys. Rev. Lett. 23, 17 (1969); Lifshitz (1964); Donsker & Varadhan (1975), Balagurov & Vaks (1974). — Iris, fire 95.
